Research On Large-scene Surveillance Video Image Stitching Technology For Traffic

With the rapid development of road traffic in China,intelligent transportation systems are receiving more and more national attention,and nowadays video surveillance systems for public transportation are popular in large numbers,but there are mutual constraints between the high resolution and wide viewing angle of image acquisition equipment,and it is impossible to obtain panoramic images of traffic congestion,traffic accident determination,and stations that take into account wide viewing angle and high resolution at the same time.This paper focuses on realizing multi-view panoramic view of video surveillance system,increasing the monitoring view of single image,and providing accurate information and efficient support for fields such as traffic management and traffic infrastructure maintenance.Therefore,it becomes extremely important to use image stitching technology to quickly realize panoramic images with wide viewing angles and high resolution.In this paper,the image stitching technology is studied through multiple cameras for image acquisition,and the research content as well as innovative work is as follows:1.Selection of feature extraction algorithms for large traffic scene video images.In order to obtain the feature points of large traffic scene video images with high matching accuracy and low reprojection error,this thesis firstly conducts an in-depth study of three most widely used and effective feature extraction algorithms,and uses experimental means to compare and analyze the characteristics of SIFT,SURF and ORB feature extraction algorithms and the effect of different scene applications,and finally selects The SIFT algorithm with high feature point matching accuracy and low reprojection error are selected as the feature extraction method in this thesis.2.Fast registration algorithm based on feature extraction of overlapping similar regions.Aiming at the problems of slow feature extraction and a large number of redundant feature points in the extracted feature points in image registration,the idea of searching the overlapping similar regions of the stitched image before feature extraction is proposed to achieve fast feature extraction and removal of redundant feature points,and the experiments prove that the method improves the feature extraction efficiency by 23.94%,the matching accuracy by 9.49%,and the completion of The efficiency of initial matching The efficiency of completing primary matching is improved by more than 47.06%.Meanwhile,for the problem of poor stability and timeliness of the Random Sample Consensus(RANSAC)algorithm,the RANSAC algorithm based on the joint sub-region sampling double-threshold constraint is proposed,which uses sub-region sampling to improve stability as well as the inner loop method of inner point random sampling double-threshold constraint,and terminates the number of iterations to improve timeliness.It is experimentally demonstrated that the algorithm improves stability while achieving an average efficiency of 61.93% over the RANSAC algorithm at the same registration accuracy on a selfbuilt traffic video image dataset.3.Image stitching based on the improved optimal stitching line.For the ghost phenomenon in the overlapping area in image stitching,the method based on the best seam-line is chosen to realize image stitching,and the improved image stitching based on the best seam-line of the graph-cut method is proposed for the situation that the seam-line passes through buildings or vehicles in image stitching,which may lead to the misalignment of the stitching effect.The improved smoothing term calculation method uses the HLS color space to construct the color difference of the smoothing term in the graph-cut method,and gives the weights of different channels to realize the importance of different channels,and then introduces the gradient difference information to make the suture line search path avoid the targets such as buildings and vehicles.Experiments prove that the improved algorithm can effectively avoid targets such as vehicles and buildings,and reduce the misalignment,unnatural color transition and ghosting in the suture images.